the labor market consequences of electricity adoption in...

The labor market consequences of electricity adoption in the

concrete industry during the Great Depression

Miguel Barroso Morin⇤

Preliminary and incomplete. Please do not circulate and do not cite without permission.

This version: August 8, 2014.

The 1920s and 1930s witnessed changes in the US labor market, with a shift away from dexterity-

intensive occupations, a productivity speedup, and low job creation. This paper asks whether

the adoption of electricity can explain these changes. The identification strategy uses a state’s

initial loading on the technology to generate electricity—hydroelectric power or coal power—as

an instrument for changes in the price of electricity. It also uses a newly digitized dataset

for the concrete industry from 1929 to 1935 to provide plant-level measures of labor market

outcomes. Technical progress in electric utilities caused, in the downstream industry of concrete,

a decrease in employment and in the labor share of income, as well as an increase in labor

quantity productivity and electrical intensity.

Keywords: electricity, Great Depression, labor market, unemployment, labor productivity, la-

bor share of income.

JEL codes: N12, N32, N62.⇤University of Cambridge (UK). Contact: [email protected] and http://www.columbia.edu/~mm3509/. This

paper is an updated version of a chapter in my PhD dissertation. I am greatly indebted to Columbia Univer-sity faculty: Suresh Naidu, Ricardo Reis, Bernard Salanié, and Jón Steinsson. I also thank Charles Calomiris,Christopher Conlon, Michael Johannes, Stephanie Schmitt-Grohé, Miguel Urquiola, and Martín Uribe for theircontinued help. Vasco Carvalho, Nicolas Ziebarth and my colleagues Audinga Baltrunaite, Christopher Boone,Jonathan Dingel, Lucie Gadenne, and Ilton Soares provided helpful comments. I gratefully acknowledge Fundaçãopara a Ciência e Tecnologia for the financial support of my doctoral dissertation.

1 Introduction

A previous paper (chapter 1 of Morin, 2014) explained US labor market changes since the 1980s

with the adoption of computers: structural changes such as a shift away from routine and

automated occupations, a productivity speed-up, and a decline in the labor share of income.

Similar changes happened in the 1920s and 1930s: a shift away from dexterity-intensive, repetitive

occupations (Gray, 2013) an increase in the growth rate of labor productivity (Field, 2003), and

low job creation (see the literature review at the end of the introduction).

This paper uses that model, designed to explain labor market changes with the adoption of

computers since the 1980s, and asks whether it can also explain labor market changes during

the 1930s. The model is general enough to apply to other technologies: this paper simply

replaces “computers” with “electricity” and the 1980s with the 1930s. Testing the model in the

context of electricity has several advantages compared to computers: electricity prices vary across

regions depending on the source of power (hydroelectric or coal) but computers prices are the

same everywhere; electricity is a homogenous good requiring no hedonic price adjustments; and

electricity is measured with consumption instead of initial investment. This test also disentangles

technology from competing explanations for labor market changes in the 1980s, such as offshoring

(Elsby, Hobijn and Sahin, 2013) and unionization (Berger, 2012): in the 1930s, offshoring was

infeasible and unionization rates were increasing (Farber and Western, 2000).

The model has two main assumptions, which find support in the economic history of electrifi-

cation. First the model assumes a decrease in the price of electricity, estimated at 7% per year

between 1899 and 1947 by Gordon (1992, Table 1) and at 5.8% between 1902 and 1950 by the

Historical Statistics of the United States (see Appendix C.3). Second, the model assumes sub-

stitutability between electrical machinery and some types of jobs: Goldin and Katz (2010, page

112) cite the example of laborers on the factory floor who were replaced by the conveyor belt,

while Jerome (1934) documents the introduction of labor-saving machinery in many industries.

To emphasize the parallels with the recent period, this paper also labels these jobs as routine

(even though a routine occupation relative to electricity is different from a routine occupation

relative to computers).

2

As an overview of the medium-term implications, the model matches the structural changes in the

1930s with these two assumptions. As electrical machinery becomes more competitive compared

to workers, firms replace one with the other. The trend of automation causes employment to shift

away from routine occupations, which substitute technology, and into nonroutine occupations,

which complement technology. The same decrease in the price of electricity has a larger effect on

the growth rate of labor productivity when electricity is cheap—because firms replace workers in

routine occupations with electrical machinery—than when electricity is expensive—because firms

forego investment in electrical machinery and hire workers in routine occupations instead. The

price of electricity has a level effect: the same decrease in the price from a lower level causes a

higher increase in the growth rate of labor productivity, which explains the productivity speedup

of the 1920s and 1930s.

As an overview of the short-term implications, the model matches the cyclical changes with the

additional assumption of labor market frictions. Firms know that they will have to hire more

nonroutine jobs in the medium-term. If they destroy nonroutine jobs during the recession, they

know that they will have to hire them back in the recovery and pay a hiring cost. To avoid

the hiring costs, firms hoard nonroutine jobs during the recession and the burden of adjustment

falls on routine jobs. Routine jobs do not entail this hiring cost in the recovery because of their

declining trend. Firms did not lay off workers in nonroutine occupations during the recession, so

they do not hire them back in the recovery. They may hire back workers in routine occupations

but, since the medium-term trend of employment in routine occupations is decreasing, routine

jobs do not recover back to peak. Total employment is constant, even as output recovers, which

is the definition of a jobless recovery.

The crucial assumption underlying this behavior of the model is the substitutability between

routine jobs and electrical machinery. If electrical machinery is equally substitutable to routine

and nonroutine jobs, as with a Cobb-Douglas production function, then the model predicts a

constant trend for the routine share of employment, the labor share of income, and productivity

growth. When a business cycle shock vanishes, the economy returns to the constant trend, so

recessions and recoveries have the same dynamics independently of the price of electricity.

This paper tests several predictions of the model. First, it uses the labor share of income to

3

test the crucial assumption of substitutability between routine jobs and electrical machinery.

If the elasticity of substitution between electrical machinery and routine jobs is greater than

1, the labor share of income should decrease as electricity becomes cheaper; if the elasticity of

substitution between electrical machinery and all jobs equals 1—as in a Cobb-Douglas production

function—the labor share of income should be unrelated to the price of electricity. Second,

the paper tests the other predictions of the model for employment, productivity, and electrical

intensity.

The ideal test of the model would be a random assignment of input prices across regions and a

subsequent analysis of the labor market outcomes. Compared to this ideal test, the first part

of the identification strategy uses geography as an instrument for the change in the price of

electricity in the 1930s. Electricity at this time came either from hydroelectric power or coal

power. Hydroelectric power had high efficiency in 1930, extracting 90% of the potential energy

of falling water, and had few opportunities for cost savings. Coal power had low efficiency,

extracting 25% of the thermal energy of coal, and had many opportunities for cost savings.1

The price of electricity decreased in regions with coal power, such as New Jersey, but not in

regions with hydroelectric power, such as California. A state’s initial loading on coal power is

an instrument for the supply-side change in the price of electricity.

The second part of the identification strategy consists of choosing the concrete industry, whose

location decisions are orthogonal to the geography of electricity prices. Concrete plants produce

a non-traded good and locate near their customers rather than near cheap electricity. The

industry has high transport costs (ready-mix concrete, for example, has to be conveyed to the

final location in a few hours) and is among the most dispersed and non-traded industries with a

Gini concentration coefficient of 30% in 1935.2 Concrete plants locate in New Jersey or California1National Electric Light Association (1931, page 43).2The most dispersed of all industries is ice cream with a Gini concentration coefficient of 17%. This coefficient

(Holmes and Stevens, 2004, page 2810) measures the difference between the distribution of economic activitycompared to population. Denote the number of states with N , the share of population in state k as popk, andthe share of activity (number of plants or total employment) state k for industry i as actki. Define the locationquotient LQ as the ratio of activity to population: LQki = actki/popk. Order the share of activity by non-decreasing order of location quotients: LQ1i LQ2i · · · LQNi. The Gini inequality coefficient for industry i

is:

Ginii = 1�NX

k=1

popk ⇥ actki + 2

NX

l=k+1

actli

!.

4

to be close to their customers, after which they react to the change in the price of electricity in

each state. Measurements of labor market outcomes for the concrete industry provide a quasi-

experiment to assess the causal effect of technical progress in electric utilities on downstream

industries.

This paper uses the universe of concrete plants from the Census of Manufactures, from 1929 to

1935, digitized for the first time for this project. It has information on employment, wage-bill,

revenue, cost of electricity, consumption of electricity, and the number and horsepower of electric

motors. Linking plants across years produces a panel of 629 continuing plants.3

The instrumental variable regressions suggest that technical progress in the electric utility indus-

try caused a decline in the labor share of income of the concrete industry and an increase in the

use of electric motors, consistent with the mechanism of capital-labor substitution in the model.

As a reminder, a Cobb-Douglas production function has constant factor shares: a decrease in the

price of an input leaves the other input shares unaffected. The empirical result in this paper is

consistent only with a production function where the elasticity of substitution between electricity

and labor is greater than 1—the crucial assumption of the theoretical model. The instrumental

variable regressions also suggest that cheaper electricity caused a decrease in employment, an

increase in labor quantity productivity, and an increase in electrical intensity. These results are

robust to several alternative specifications, such as dropping plants located near the construction

of dams. Furthermore, the coal instrument is not correlated with banking failures or housing

construction over the 1920s, lending support to the validity of the instrument.

Related literature. This paper relates to several strands of the literature: electrification during

the 1930s, the parallels between electricity and computers, and the jobless recovery from the

Great Depression. On electrification in the 1930s, several studies have used aggregate-level data

or Ordinary Least Squares to assess the effects of electrification on the labor market. Gray (2013)

studied worker-level evidence from the first half of the 20th century and found that electrification

was correlated with a shift away from occupations intensive in dexterity skills, similar to the3A previous version of this paper used less precise linking of plants across years and a sample size of 742 plants.

This version uses automated linking software with a bigram comparator, more precise linking across years, and asmaller sample size of 629 plants. The quantitative results are similar but the standard errors are smaller becauseplants less likely to be the same in the first version added noise to the regression.

5

findings of Autor, Levy and Murnane (2003) for computerization in the late 20th century. Field

(2003) used aggregate-level growth accounting and argued that the 1930s had an unprecedented

increase in TFP and were the “most technologically progressive decade of the century” because of

electricity. Woolf (1984) used industry-level data from the Census of Manufactures between 1909

and 1929 and found that “firms sought labor-saving and capital-using techniques in response to

cheaper energy ... [and reduced] labor’s share of income.” The evidence from previous studies is

consistent with the thesis of this paper, whose contribution is to use plant-level data, to propose

a new instrument for the adoption of electricity, and to test broader implications of technology

adoption (labor share of income, employment, and productivity).

This paper also relates to the literature on the parallels between electricity and computers. David

(1990) argued that both electricity and computers generated productivity growth in the wider

economy after a long lag, causing the productivity speedups of the 1920s and 1990s. Syverson

(2013) found that the speedup in labor productivity of the 1990s was of a similar magnitude as

that of the 1920s, documented by Kendrick (1961, page 71).4

This paper also relates to the literature on the jobless recovery and technological unemployment

during the Great Depression. Irving Fisher in 1928 proposed technology as an explanation for the

jobless recovery from the 1927 recession: “increased productivity per worker, aided by improved

machinery and organization and more willing labor, is partly responsible for the anomaly of grow-

ing unemployment during an extended period of increased business activity” (quoted by Woirol,

1996, page 28). Keynes coined the term of “technological unemployment”: “unemployment due

to our discovery of means of economising the use of labour outrunning the pace at which we

can find new uses for labour.” Frances Perkins, secretary of the Department of Labor, stated in

a Congressional testimony in 1935 that “you would be surprised at the number of labor-saving

devices which have been introduced in industry in the last 2 or 3 years” (Committee on Finance,

1935, page 206). The New York Times invented the expression “jobless recovery” in the 1930s:

“During November [of 1938, the Works Progress Administration] rolls showed some decline, but4Field (2011, page 25) questioned the exact dating of the productivity speedup of Kendrick because of his

choice of dates: “The problem is that Kendrick compared a fully employed economy in 1929 with a 1937 economyin which 14.3 percent of the labor force was still out of work ... If we seek a peacetime peak-to-peak comparison,we are better served by choosing as an endpoint 1941, when unemployment, although still averaging 9.9 percent,was closer to what it was in 1929 but before war spending or production could seriously have influenced theeconomy.”

6

it was slight enough to make observers wonder whether the country were experiencing a ’jobless

recovery.” ’5 Relative to this literature, the contribution of this paper is to suggest the decline in

the price of electricity as a possible reason for technological unemployment.

2 Data and definitions

This paper assesses the effect of technical progress in electric utilities on labor market variables.

It uses two data sources at the state-level from publications by the Census Bureau and at the

plant-level from micro-data at the National Archives. The Census Bureau published a state-level

summary of the electric light and power industry in 1927 and 1937. It also published state-level

information on other variables, such as wages in manufacturing in 1929 and 1935 in the state-

and industry-level publications of the Census of Manufactures.

The plant-level dataset is from the Census of Manufactures in 1929 and 1935, which covers the

universe of manufacturing plants with sales above five thousand dollars.6 This dataset is at

the National Archives and Records Administration in Washington D.C. Two barriers prevent

the wider use of this dataset: the schedules are in paper or microfilm format and the National

Archives protect them with in-house access only. This paper focuses on the concrete industry,

digitized for the first time for this project. I scanned all the microfilm schedules (around 2,500

for 1929 and 1,100 for 1935). The archivists marked as lost one microfilm roll with 300 plants in

1935 for states Alabama to Iowa but I was able to locate a backup copy in a different location.

No schedules from the Census of Manufactures are missing from my sample. A professional

data entry firm tabulated these schedules into electronic format. I verified the tabulations and

corrected outliers, such as missing commas in the separation of cents and dollars. I also cleaned

the names of states, counties and cities. The Census Bureau had no unique plant identifier and

I matched the plants across years based on their name, location and ownership (see Appendix

2). From the 3,500 plants present in both 1929 and 1935, I obtained a panel of 629 continuing

plants.5Article “Jobless recovery?” of 27 November 1938.6This threshold in 1929 corresponds to around $66 thousand today and is high above the average sales for the

concrete industry of $38 thousand in 1929 prices.

7

The concrete industry has three advantages for identification. First, it sells non-traded products

(Syverson, 2004), which guarantees that concrete plants locate near their customers and their

geographic distribution is exogenous to the regional variation in the price of electricity. Second,

the concrete industry is intensive in electricity: continuing plants spent on average 1% of revenue

in electricity in 1929, which puts concrete in the upper sixth of manufacturing industries that

use the most electricity. Third, concrete plants are small and bought all of their electricity from

the grid: the Census Bureau asked about generation of electricity, which is zero for all firms in

the balanced panel.

The Census asked about production by quantity and value, employment, wages, number of

electric motors, horsepower of electric motors, kilowatt-hours purchased and their cost, and

kilowatt-hours generated. The top panel of Table 1 shows summary statistics for continuing

concrete plants. The concrete industry has many small plants, with an average of 21 employees.

The bottom panel shows summary statistics for the change between 1929 and 1935. On average,

concrete plants had a decrease in output, the labor share, employment, the price of electricity,

and an increase in the horsepower of electric motors.

Concrete plants use labor-saving electrical machinery at several stages of production of concrete:

machinery for crushing and grinding stones into a finer aggregate, machinery for pumping and

unloading units to convey cement, electric power shovels and conveyor belts or elevators to move

materials, mixing machines that produce a more homogenous product with less cement compared

to manual mixing, and waste-heat boilers (Jerome, 1934, page 80; Orchard, 1962, page 404).

The concrete industry had a decline in the labor share of revenue of 14 percentage points, from

28.7% in 1909 to 14.4% in 1939, illustrated in Figure 1. Half of this decrease, or 7 percentage

points, occurred during the Great Depression. The other half occurred during the other recessions

of 1927 and 1937. 7

7The labor share of value added shows similar numbers: a decline of 17 percentage points from 46% in 1909to 29% in 1939, of which 7 percentage points occurred during the Great Depression. Nevertheless, this measureis less comparable across years because it sometimes omits fuel and energy.

8

Summary statistics for 1929

Number Employment Average Electricity Electricity and Kilowatt

of of all employment share fuel share hours

plants plants per plant of income of income purchased

629 7964 21 1.3% 2.4% 18,076

Summary statistics for the change between 1929 and 1935

Change from 1929 to 1935 (log-points, annualized) Mean S.d.

Output value -0.09 0.15

Labor share -0.02 0.10

Employment -0.04 0.13

State-level cost of electricity -0.02 0.01

Horsepower of electric motors 0.01 0.13

Table 1: Summary statistics for the concrete industry.

1520

2530

Labo

r sha

re o

f rev

enue

(%)

1909 1929 1933 1939Year

Figure 1: The decline in the labor share of revenue of the concrete industry accelerated duringthe Great Depression.Details: wages divided by revenue every two years from 1909 to 1939, from the publication Census of Manufacturesfor the year 1939. Shaded areas are NBER recessions.

9

3 Methodology

3.1 Overview of the model

This subsection summarizes the production side of the theoretical model. The General Equilib-

rium properties of the model (household side and equilibrium of the labor, product, and capital

markets) are omitted here and the interested reader is referred to Morin (2014). Plant i rents

two types of capital, electric capital KE,i,t and non-electric capital KNE,i,t. The first assumption

is a long-term decrease in the rental rate of electric capital.

Assumption 1. The rental rate rE,i,t of electric capital decreases exogenously with time:

rE,i,t & in t.

Plant i hires workers in two types of occupations, routine occupations LR,i,t and nonroutine

occupations LNR,i,t . The production function of plant i is:

Yi,t = Ai,t K↵NE,i,t L

�NR,i,tM

�i,t,

Mt =

✓K

��1�

E,i,t + L

��1�

R,i,t

◆ ��1

, (1)

where Ai,t is Total Factor productivity. The production function has constant returns to scale,

with ↵+�+� = 1. This production function has Cobb-Douglas aggregation of three factors: non-

electric capital KNE,i,t, employment in nonroutine occupations LNR,i,t, and a third factor, which

is a Constant-Elasticity-of-Substitution aggregate between electric capital KE,i,t and employment

in routine occupations LR,i,t. The second crucial assumption is gross substitutability of electric

capital and employment in routine occupations tasks in the production function:

Assumption 2. The elasticity of substitution between electric capital and employment in routine

occupations is greater or equal to 1:

� � 1.

10

Plant i operates under perfect competition and has profits

profitsi,t = pi,tYi,t � wi,t (LNR,i,t + LR,i,t)� rNE,i,tKNE,i,t � rE,i,tKE,i,t,

where wi,t is the wage, pi,t is the price of output. The firm maximizes the present value of profits,

discounted with the market interest rate rt.

As in Morin (2014, ch 1), the wage wi,t is the same for routine and nonroutine occupations

because the household supplies both tasks with no friction. Unlike Morin (2014, ch 1), hiring

costs are zero in this setting, which guarantees a closed-form solution. The interested reader is

referred to Morin (2014, ch 1) for more details on the supply side of the labor market and the

short-term implications of the model when including adjustment costs.

3.2 Testable predictions

Morin (2014, ch 1) shows that the General Equilibrium model with hiring costs has five predic-

tions for the labor market: (1) labor productivity speeds up, (2) employment shifts away from

routine occupations and into nonroutine occupations, (3) the labor share of income declines, (4)

recessions accelerate the structural decline in routine occupations, (5) recoveries from recessions

are jobless, i.e. employment recovers slower than output.

Testing the theory requires choosing the predictions to test with the available data. One predic-

tion of the model—routinization of production—is the subject of Gray (2013). She merged the

worker-level Census of Population from 1900 to 1950 to the Dictionary of Occupational Titles.

She defined an occupation as routine if it required high dexterity and low manual or clerical skills.

She found that states with faster electrification also shifted away from these routine occupations,

similar to the findings of Autor, Levy and Murnane (2003) for computerization in the late 20th

century and consistent with the thesis in this paper.

The prediction of acceleration of routinization cannot be tested with this dataset. This prediction

requires high-frequency information on employment by detailed occupations, which is unavail-

able in the Census of Manufactures (see Appendix A.3). If electricity complements nonroutine

11

occupations, the positive impact on nonroutine jobs could offset the negative impact on routine

jobs, which biases against finding a positive net effect.

Therefore, the rest of this paper focuses on the three testable predictions: employment, productivity,

and the labor share of income. The prediction of jobless recoveries finds some support in the

data, with electricity adoption causing a slower recovery of employment between 1933 and 1935.

A more general formulation of this prediction, for example if labor supply has a reallocation cost

across occupations, is that electricity causes a decrease in employment over the whole period of

1929-1935, which is also tested.

The period covered is 1929 and 1935 for three reasons. First, the plant schedules of the Census of

Manufactures survived only for this period and the years before or after were destroyed. Access

to plant-level data is important in order to link plants across years and avoid compositional

bias due to the turnover of plants. It also contains more information, such as output in tons of

concrete and the horsepower of electric motors, which is not otherwise available. Second, the

major turmoil in labor markets during the Great Depression provides variation in the dependent

variables and allows a more precise estimation of the regression coefficients.

3.3 Linear regressions

The model implies the following non-linear equations for the labor share of income and the

electric capital-labor ratio (see Appendix B for proofs):

wj,tLj,t

pj,tYj,t= � + �

1 +

✓rE,j,t

wj,t

◆1��!�1

, (2)

KE,j,t

Lj,t=

✓rE,j,t

wj,t

◆�1 �

�

+

✓1 +

�

�

◆✓rE,j,t

wj,t

◆��1!�1

, (3)

where j indexes a unit of observation such as firms i or regions k. The equation for labor quantity

productivity Yj,t/Lj,t is similar and omitted.

These equations include a “level effect:” when electrical machinery is too expensive and � > 1,

the non-linear term vanishes from the equation and the labor share of income is constant. When

12

technology is too expensive, a decrease in the price of electricity has little impact on the economy,

as firms prefer to hire workers instead. The “level effect” was discussed in Morin (2014, ch 1).

To translate these predictions into regression equations, I consider the log-linear version of the

equations. The problems in using a log-linear version of non-linear equations with a level effect

are minimal because the level effect is more important over decades of decrease in the price of the

technology, rather than the six years between 1929 and 1935. The next simulations illustrate that

the log-linear regressions are an accurate approximation to the non-linear relationships, I use the

General Equilibrium model from the Morin (2014, ch 1) to simulate 300 artificial economies. All

simulations use the same parameters, equal to those from the calibration in Morin (2014, ch 1),

except for the rate of decrease � in the price of electricity, which follows a uniform distribution

between 1% and 18%. I solve the model for each artificial economy j and estimate the following

log-linear equations:

� log

wj,tLj,t

pj,tYj,t= 0.0011 + 0.047� log

✓rE,j,t

wj,t

◆+ error (4)

� log

KE,j,t

Lj,t= �0.0023� 1.556� log

✓rE,j,t

wj,t

◆+ error (5)

Under the assumption � > 1, the slope coefficient should be positive for the labor share of

income and smaller than -1 for the computer capital-labor ratio. The scatter plot in Figure 2

shows that the log-linear regression from the model is an accurate approximation to the non-linear

expression.

Two further difficulties arise in the context of electricity. First, the relative rental rate rE,j,t/wj,t

of electrical machinery is unobserved and I use the price of electricity in cents per kilowatt-hour

as a proxy, which implies measurement error and an attenuation bias toward zero.8 Second, the

average price of electricity at the plant-level is far from the marginal price: several forms of fixed

costs (see Appendix C) introduce measurement error in the price of electricity paid by small8The usage cost of electricity has two components: the price of electricity in kilowatt-hours and the rental

rate of an electric motor. Regional variation in the usage costs stems mostly from the price of electricity becausethe rental rate of electric motors is likely to be the same for all regions. The rental rate of an electric motor hasthree components: the interest rate, the price of investment, and the depreciation rate. Each of these componentsshould have similar values across regions: the interest rate was set by the Federal Reserve for all regions andthe electrical machinery industry was concentrated in five states which served a national market with similarinvestment prices and depreciation rates.

13

concrete plants. Fixed costs should lose importance when considering a larger entity such as the

state, whose average price of electricity should be closer to the marginal price. The preferred

measure of the price of electricity is the state-level average price from the Census of Electric

Light and Power Stations for 1927 and 1937.9 This measure minimizes the importance of fixed

costs, making the average price closer to marginal price, and is close to the price of electricity

paid by industrial users, since power stations sold on average 69% of their current to industrial

consumers.10

The regression equations are:

� log

wi,tLi,t

pi,tYi,t= constant+ a� log (pE,k,t) + error. (6)

� log

Yi,t

Li,t= constant+ b� log (pE,k,t) + error (7)

� logLi,t = constant+ c� log (pE,k,t) + error (8)

� log

KE,i,t

Li,t= constant+ d� log (pE,k,t) + error (9)

where i indexes plants, k indexes states, wi,tLi,t is the aggregate wage-bill at the plant-level,

pi,tYi,t is the output value at the plant-level, pE,k,t is the change in the price of electricity at the

state-level, Yi,t/Li,t is labor quantity productivity in tons of concrete, Li,t is employment, and

KE,i,t/Li,t is a measure of electrical intensity at the plant-level (the horsepower of electric motors

per worker). Alternative regressions use on the right-hand side � log (pE,k,t/wk,t), the change in

the price of electricity relative to the state-level wage: if so, the left-hand side of (6) uses the

wage at the plant-level for the concrete industry and the right-hand side uses the wage at the

state-level for all manufacturing industries. The theory predicts a > 0, b < 0, c > 0, and d < �1:

a decrease in the price of electricity should cause a decrease in the labor share of income, an

increase in productivity, a decrease in employment, and an increase in electrical intensity.11

9Stigler and Friedland (1962) used this measure to assess the effect of regulation on electricity prices. To thebest of my knowledge, the Census of Electric Light and Power Stations is the only source of data for the price ofelectricity at the state-level during this period.

10Census of Electric Light and Power Stations, 1927, page 51.11The model normalizes the price of output pi,t to 1, so other prices are in real terms. The regressions use a

nominal price with no deflator—deflating prices by a nation-wide price or wage index would affect the interceptof the regression and not the slope.

14

3.4 Endogeneity and an instrument

The identifying assumptions for regression equations (6) to (9) are that the average price of

electricity, of labor, and of output are close to the marginal prices and that the error term is

uncorrelated with the regressors. Then a and a

0 are consistent and unbiased estimators.

Estimating a regression of quantities on prices raises concerns about endogeneity and is a chal-

lenge to identification: it is unclear whether the regression estimates the demand or supply

equation. This paper is interested in the demand for electricity and requires an instrument that

shifts the electricity supply curve and not the demand curve. This endogeneity should bias the

estimation of the downward-sloping electricity demand curve toward the upward-sloping electric-

ity supply curve. The coefficients should be further away from zero in Instrumental Variables

(IV) compared to Ordinary Least Squares (OLS). A similar argument suggests that endogeneity

also biases the coefficient on the labor share of income toward zero because the labor share of

income is decreasing in the electric capital-labor ratio in the model.

The identification strategy to deal with the endogeneity bias consists of two parts: using geog-

raphy as an instrument for the change in the price of electricity and choosing the non-traded

industry of concrete. As an instrument for the supply-side change in the price of electricity,

this paper uses the share of coal in the generation of electricity in 1927. In 1930, power plants

extracted 90% of the potential energy of falling water and had few opportunities for cost-saving

innovations. Power plants extracted 25% of the potential energy of burning coal to power steam

turbines, had many opportunities for cost-saving innovations.12 The generation of electricity

from coal improved thanks to a “rise in steam pressures and steam temperatures used, and ...

the experimental introduction of a second working fluid in an independent cycle supplementing

that of the steam.”13 These innovations increased the thermal efficiency of fuel: “In 1928, the

same amount of energy was produced with 71 per cent less fuel than would have been required

in 1904.”14

Technical progress in the generation of electricity from coal impacted regions differently depend-12National Electric Light Association (1931, page 43).13Census of Electric Light and Power Stations (1927, page 82)14Electrical Research Statistics (1929). See also Sleight (1930, page 57) for a similar finding.

15

ing on their initial dependence on this technology. Regions with access to hydroelectric power,

such as Minnesota or California, have cheap electricity but the price of electricity is roughly con-

stant. Regions without hydroelectric power, such as North Dakota or New Jersey, have initially

more expensive electricity but the price of electricity decreases. Figure 3 illustrates the pattern

of convergence across states. Figure 4 shows the first-stage of the instrument at the state-level:

states with initially larger dependence on coal power also had a decrease in the relative price of

electricity. The two measures of the price of electricity in this paper are:

� log (pE,k,t) =1

10

log

✓pE,k,1937

pE,k,1927

◆, � log

✓pE,k,t

wk,t

◆=

1

10

log

✓pE,k,1937

pE,k,1927

◆�1

6

log

✓wk,1935

wk,1929

◆,

where the price of electricity is the average price of electricity for ultimate consumers from the

Census of Electric Light and Power Stations in 1927 and 1937 and the wage is the industry-wide

average wage for wage-earners and salaried workers for all manufacturing firms in 1929 and 1935.

Four arguments support the validity of this instrument. First, concrete plants do not sort

geographically depending on the price of electricity: the concrete industry sells a non-traded

product and locates near its customers. Second, the instrument should affect electric utilities

on the supply side of the electricity market but not concrete plants on the demand side of the

market. Third, the instrument is an initial level and the outcome variables are changes. Omitted

variables in levels, such as the skill composition of the workforce or the density of the road

network, are differenced out in the regressions. Fourth, using ratios at the plant-level, such as

labor productivity or the labor share of income, implies the absence of that plant-level shocks

that affect the numerator and denominator similarly, such as TFP shocks.

A possible violation of the exclusion restriction concerns omitted variables that change through

time. For example, cities high in hydroelectric power may attract more government programs

for dam construction, which would increase demand for concrete in regions with hydroelectric

power compared to regions with coal power. This increase in demand may be met with the more

adjustable factors, such as labor or materials. To address this concern, I run a falsification test

with the materials share of income and also run a robustness test by dropping states with dam

construction from the sample.

16

-.006

-.004

-.002

0.0

02Ch

ange

in la

bor s

hare

of i

ncom

e

-.15 -.1 -.05 0Change in relative price of electricity

Slope: 0.05 t-statistic: 104.97R2: 0.97 Correlation: 99% Observations: 300

Figure 2: The non-linear relationship in the model is close to a linear one for short periods oftime.The vertical line corresponds to a decrease in the price of electricity of � = 7%.

AL

AR

CACO

CTFLGA

IA

IDIL

INKS

KY

LA

MAMD

ME

MI MO

NC

NENHNJ

NY

OH

OK

ORPA RI

SC

SD

TN

TX

UT

VAWA WI

WV

MN

ND-.6-.4

-.20

.2Ch

ange

in p

rice,

192

7-19

37

0 .5 1 1.5 2Logarithm of price of electricity in 1927

Slope: -0.30 t-statistic: -8.50R2: 0.66 Correlation: -81% Observations: 40

Figure 3: The price of electricity converged across states between 1927 and 1937.The two neighboring states of Minnesota and North Dakota have a different color and a larger font. The outlierstates of Mississippi and Arizona are omitted.

17

4 Results

This section presents the evidence for � > 1 and for the causal link between electricity and labor

market outcomes: the labor share of income, labor productivity, and employment. Concrete

plants with access to cheaper electricity also reduce their labor share of income, increase labor

quantity productivity, reduce employment, and have a higher electric capital-labor ratio. The

results are robust to several alternative specifications and the coal instrument is not correlated

to banking failures or to the growth in construction over the 1920s.

4.1 Baseline results

Table 2 shows the baseline results in instrumental variables and Table 3 shows the results in

reduced-form. The exogenous decrease in the price of electricity caused a decrease in the labor

share of revenue, an increase in labor quantity productivity, a decrease in employment, and an

increase in electrical intensity. The regression of labor quantity productivity (the number of

concrete tons divided by employment) suggests that the results are not due to deflation or other

price channels. The regression of the labor share of revenue supports the crucial assumption in

the model. The coefficient for the labor share is proportional to ��1: it should be positive under

the assumption � > 1 and zero under � = 1. The IV regression of electrical intensity traces

the demand curve and finds a negative coefficient: cheaper electricity induces more horsepower

per worker. The theory predicts that the coefficients on electric capital-labor ratios should be

smaller than -1 and the regressions confirm that prediction. The coefficients are economically

and statistically significant. The standard errors in all plant-level regressions are clustered at

the state-level and all variables are “winsorized” at the 1% level.

A back-of-the-envelope calculation with the point estimates in reduced form suggest that tech-

nical progress in the coal technology may explain the decrease in the labor share of revenue of

the concrete industry: it decreased on average 1.6% per year, while the reduced-form regression

predicts a decrease of 1.9% (equal to the reduced-form coefficient of 0.0247 times the average coal

dependence in the sample of 0.78). That is, technical progress over-predicts the decrease in the

18

AL

AR

AZ

CA

CO

CT DCDE

FLGA

IA

ID

IL

IN

KS

KY

LA

MAMD

ME

MIMN MO

NC

NDNENH

NJ

NY

OH

OK

OR

PA RI

SC

SD

TN

TX

VAWAWI

WV-.04

-.03

-.02

-.01

0Ch

ange

in p

rice

of e

lect

ricity

(27-

37)

.2 .4 .6 .8 1Coal share of power in 1927

Slope: -0.02 t-statistic: -4.38 F-statistic: 19.2R2: 0.32 Observations: 42

Figure 4: First-stage regression: an initially higher share of coal in power generation in 1927causes a subsequent decrease in the relative price of electricity.Larger circles represent states with more plants but the regression has the same weight for all states.

Dependent variable: � log

wi,tLi,t

pi,tYi,t� log

Yi,t

Li,t� log (Li,t) � log

KE,i,t

Li,t

� log (pE,k,t) 1.954** -4.960* 3.452*** -6.945***

(state-level, instrumented) (0.965) (2.904) (1.311) (2.276)

Constant 0.0285 -0.156** 0.0357 -0.0915*

(0.0225) (0.0661) (0.0280) (0.0518)

Observations 620 551 620 474

Number or states / clusters 42 42 42 39Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 2: Baseline results in IV: the decrease in the price of electricity caused a decrease in thelabor share of revenue, an increase in labor quantity productivity, a decrease in employment, andan increase in electrical intensity.

19

labor share of revenue for continuing plants. This thought experiment assesses the net contribu-

tion of electricity and holds constant other factors, such as wages that may have national-level

shifters such as the National Recovery Act of 1933.

To ensure that the results are in line with the theory, which predicts that the relevant variable

is the price of electricity relative to wages, Table 4 shows the IV regressions with the price of

electricity relative to wages in all of manufacturing. The coefficients have similar magnitudes

and lose some statistical significance.

4.2 Falsification tests

Tables 5 shows four falsification tests, or placebo regressions. The first two test an alternative

channel of a demand shock: if the coal instrument were correlated with a demand shock, it could

explain both a decrease in production and in employment. The first two columns of Table 5

suggest that such is not the case: the coal instrument is not a statistically significant predictor of

the change in value or quantity of output. The last two test the effect of the instrument on the

materials and fuel share of revenue. One might be concerned that, since revenue shares sum to

one, the effect of electricity on the labor share may be a arithmetic consequence of the increase

in the share of materials or fuel. The last two rows of Table 5 suggest that such is not the case

either: the IV regressions have statistically insignificant coefficients.

4.3 Geography of the coal share of power

Using state-level geography as an instrument has the drawback that the instrument corresponds

to inland regions as opposed to the coasts. Figure 5 shows that the mountains in the West and

East Coast provide the altitude differentials necessary for hydroelectric power while the Great

Plains need to use coal power. Some variation persists within region, such as the neighboring

states of North Dakota with 100% coal power versus Minnesota with 36% coal power, or the

neighboring states of Florida with 98% coal power versus Georgia with 13% coal power. Never-

theless, the within-region variation is not sufficient to confirm the baseline results. Hydroelectric

20


wi,tLi,t

pi,tYi,t� log

Yi,t


KE,i,t

Li,t

coalk,1927 -0.0247** 0.0625*** -0.0435*** 0.0935**

(state-level) (0.0121) (0.0229) (0.0141) (0.0361)

Constant 0.00345 -0.0928*** -0.00849 -0.00738

(0.00926) (0.0185) (0.0108) (0.0308)


R-squared 0.005 0.009 0.007 0.021


*** p<0.01, ** p<0.05, * p<0.1

Table 3: Baseline results in reduced-form: initially higher coal dependence caused a decrease inthe labor share of revenue, an increase in labor quantity productivity, a decrease in employment,and an increase in electrical intensity.

Dependent variable � log

wi,tLi,t

pi,tYi,t� log

Yi,t


KE,i,t

Li,t

� log (pE,k,t/wk,t,mfg) 1.320* -3.482** 2.333* -4.801*


Constant -0.0398*** 0.0199 -0.0849*** 0.153***

(0.0133) (0.0316) (0.0236) (0.0498)


Number of states / clusters 42 42 42 39Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 4: The baseline results are robust to using the relative price of electricity instead of theabsolute price.

21

power requires falling water and is close to the map of mountains in the United States, a conse-

quence of using geography as an instrument for the change in the price of electricity depending

on the source of power. If plants in the mountain regions are affected differently during the

Depression, it may invalidate the exclusion restriction of the Instrumental Variable approach.

One possibility is that mountain regions have government programs for building dams. Table

6 shows that the baseline results are robust to dropping counties within 50 miles of dams un-

der construction, giving confidence that the instrument is valid and the results are not due to

government demand for concrete products.

Several other correlations suggest that the coal instrument is not picking up alternative channels:

the instrument does not predict the previous growth in housing construction over the 1920s, it

does not predict the subsequent bank failures across states. Another threat to identification

occurs if the share of coal in electric power generation reacts to changes in electricity demand

and in aggregate demand. Figure 7 suggests that the change in coal capacity is uncorrelated

with the change in housing construction over the 1920s.

4.4 Jobless recovery

The last prediction of jobless recoveries finds some support in the data. Table 7 shows the results

for employment and productivity between 1933 and 1935, in reduced-form, for plants that report

both employment and physical output. The decrease in the price of electricity caused a slower

recovery of employment and strong productivity gains, both in IV and in reduced-form, with

a magnitude twice as large as the coefficients for 1929-1935. Nevertheless, these results should

be taken with a grain of salt: they are not entirely robust to other specifications, such as the

aggregation of all plants.

4.5 State-level results for all plants

The predictions of the model for the labor share and labor productivity are also valid at the

state-level for all plants, not just continuing plants. To run state-level regressions, I aggregate

22

Dependent variable � log (pi,tYi,t) � log (Yi,t) � log

materialsi,tpi,tYi,t

�

fueli,tpi,tYi,t

� log (pE,k,t) 1.971 -1.647 1.288 -0.162


Constant -0.0492 -0.123** 0.0316* -0.00284

(0.0359) (0.0545) (0.0181) (0.00242)


Number of states / clusters 42 42 42 42

First-stage F -statistic 19.15 19.15 19.15 19.15Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 5: Falsification tests in instrumented variables: the decrease in the price of electricity hasno effect on the materials share of revenue.As the fuel share of revenue is small, around 1%, this regression uses the percentage point change instead of thelog-change.

Figure 5: Map of the share of coal power in 1927 (darker blue implies higher coal share).

23


wi,tLi,t

pi,tYi,t� log

Yi,t


KE,i,t

Li,t

� log (pE,k,t) 1.952* -5.172* 2.587* -8.176*


Constant 0.0297 -0.165** 0.0137 -0.123

(0.0266) (0.0713) (0.0345) (0.104)



*** p<0.01, ** p<0.05, * p<0.1

Table 6: The baseline results are robust to dropping counties within 50 miles of dams under con-struction.Details: The latitude and longitude by city is from Gaslamp Media (2014), which “compiled from acity/county/state database and geocoded with Google Maps.” The list of counties with dam construction is fromHay (1991) for dams completed between 1930 and 1940. The latitude and longitude of a county with dam construc-tion is the average of all cities in that county. The closest distance from county X to a dam under constructionis the minimum Haversine distance from all cities in county X to all cities in counties with dam construction.

AL

AR

AZ

CA

CO

CTDE

FL

GA

IA

IDILIN

KSKY

LA

MA

MD

ME

MI

MN

MOMS

MT

NCND

NE

NH

NJNM

NV

NY

OH

OK

OR

PA

RI

SC

TNUT

VAVT

WA

WI

WV

0.2

.4.6

% a

ll dep

osits

sus

pend

ed, 3

0-33

0 .2 .4 .6 .8 1Coal share in 1927

Slope: -0.03 t-statistic: -0.48R2: 0.01 Observations: 45 Correlation: -7%

Figure 6: The coal instrument is unrelated to bank failures across states.Details: bank failures is the percentage of deposits of all banks in 1930 that were suspended in 1930, 1931, 1932,or 1933.

24

AL

AR AZ

CA

CO

CTDE

FL

GA

IA

IDILIN

KS

KY

LA

MAMD

MEMI

MNMOMS

MT NC

ND

NE

NHNJ

NM NV

NYOH

OK OR

PA RI

SC

SDTN

TXUT

VA

VTWA

WIWV

WY

-.4-.2

0.2

.4Ch

ange

in c

oal c

apac

ity, 1

922-

1927

-.5 0 .5Growth in housing construction, 20-24 to 25-29

Slope: -0.10 t-statistic: -1.26R2: 0.03 Observations: 48 Correlation: -18%

Figure 7: The change in coal capacity between 1922 and 1927 is uncorrelated with the growth inhousing construction from 190-1924 to 1925-1929.Details: housing is the number of dwellings built over each quinquennium (1925-1929 or 1920-1924), kindlyprovided by Kimbrough and Snowden (2007).

Dependent variable � log (Li,t) |1933�1935 � log (Yi,t/Li,t) |1933�1935

coalk,1927 -0.118** 0.213***

(state-level) (0.0506) (0.0713)

Constant 0.261*** -0.107**

(0.0448) (0.0515)

Observations 467 467

R-squared 0.009 0.014

States / number of clusters 42 42Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 7: A higher loading on the coal technology caused a weaker recovery of employment andstrong productivity gains.

25

employment, total output value, wage bill, and electric horsepower at the state-level, and then

I compute annualized changes. Table 8 suggests that the results persist for the labor share and

productivity with similar magnitudes as the plant-level regressions. The effects on employment

and horsepower per worker become statistically insignificant: this could be due to compositional

bias and the creative destruction of concrete plants, one avenue for future research.

Dependent variable � log

wk,tLk,t

pk,tYk,t� log

Yk,t

Lk,t� log (Lk,t) � log

KE,k,t

Lk,t

� log (pE,k,t/wk,t,mfg) 1.141** -3.371** 1.103 -2.548


Constant -0.0448*** 0.104*** -0.150*** 0.110***

(0.0111) (0.0403) (0.0205) (0.0420)

Observations 46 45 46 44Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 8: State-level results in IV: the decrease in the price of electricity caused an increase inlabor productivity (value or quantity) and a decrease in the labor share of value added.Details: the regression is weighted by initial employment in each state. Output value pk,tYk,t at the state-level isvalue added instead of revenue.

26

5 Conclusion

This paper tests the model of labor market changes based on capital-labor substitution in the

context of electricity and provides two contributions. First, it uses a plant-level dataset from the

concrete industry during the 1930s, digitized for the first time for this project. This plant-level

dataset has finer detail than the Census of Manufactures state-level publications for the concrete

industry and allows a more precise test of the model, e.g. by considering continuing plants

and by excluding plants near dam construction. Second, the identification strategy uses a new

instrument—a state’s initial loading on the coal technology—to isolate the exogenous shift in the

electricity supply curve. Consistent with the predictions of the model, a decrease in the price of

electricity caused a decrease in employment, an increase in productivity, and a decrease in the

labor share of income. The results imply that the elasticity of substitution between electricity

and labor is greater than 1. Some occupations may be more replaced by electrical machinery

than others, such as the routine, dexterity-intensive occupations described by Gray (2013).

Two questions arise for future research. First, this paper used electricity prices at the state-level

and ignored within-state variation in electricity prices. As an example of within-state variation,

plants in Buffalo use hydroelectricity from Niagara Falls while plants in New York City use coal

electricity. Further data digitization from the Census of Manufactures could produce electricity

prices for industrial consumers at the county or city level, which would allow researchers to

leverage the plant-level variation of this dataset. Second, the model’s predictions for employment

find support for continuing plants but not for all plants. An investigation of the characteristics of

entry and exit in the concrete industry would provide a more accurate description of aggregate-

level employment during the Great Depression. These questions are left for future research.

It is a surprising conclusion that the model, designed to explain labor market changes with the

adoption of computers since the 1980s, can also explain labor market changes in the concrete

industry with the adoption of electricity in the 1930s. By emphasizing the parallels between

two eras of rapid technology adoption, this paper relates to the literature on general purpose

technologies throughout history: “whole eras of technical progress and growth appear to be

driven by a few ’General Purpose Technologies,’ such as the steam engine, the electric motor,

27

and semiconductors.” (Bresnahan and Trajtenberg, 1995). A recent debate has focused on the

importance of recent General Purpose Technologies compared to previous ones. Jovanovic and

Rousseau (2005) think that “electricity and information technology [are] probably are the two

most important General Purpose Technologies so far.” Gordon (1999) disagrees and suggests

that we may face decreasing returns in the invention of new technologies: “electricity . . . was a

much more profound creator of productivity growth than anything that has happened recently

. . . this was a unique event that will not be replicated in the lifetimes of our generation or that

which follows us.” It is an open question whether the next General Purpose Technology will be as

important as previous ones and whether the historical patterns of the output and labor markets

will repeat themselves.

28

References

Autor, David, Frank Levy, and Richard Murnane. 2003. “The Skill Content of Re-

cent Technological Change: An Empirical Exploration.” The Quarterly Journal of Economics,

118(4): 1279–1333.

Berger, David. 2012. “Countercyclical Restructuring and Jobless Recoveries.” Manuscript, Yale

University (accessed 7 November 2012).

Berman, Eli, John Bound, and Zvi Griliches. 1994. “Changes in the Demand for Skilled

Labor within U.S. Manufacturing: Evidence from the Annual Survey of Manufacturers.” The

Quarterly Journal of Economics, 109(2): 367–397.

Bresnahan, Timothy, and Daniel Raff. 1991. “Intra-Industry Heterogeneity and the Great

Depression: The American Motor Vehicles Industry, 1929-1935.” The Journal of Economic

History, 51(2): 317–331.

Bresnahan, Timothy, and Manuel Trajtenberg. 1995. “General Purpose Technologies:

’Engines of growth’?” Journal of Econometrics, 65: 83–108.

Committee on Finance, House of Representatives. 1935. Economic Security Act. United

States Government Printing Office.

David, Paul. 1990. “The Dynamo and the Computer: An Historical Perspective on the Modern

Productivity Paradox.” The American Economic Review, 80(2): 355–361.

Elsby, Michael, Bart Hobijn, and Aysegul Sahin. 2013. “The Decline of the US Labor

Share.” Brookings Panel on Economic Activity, Fall 2013.

Farber, Henry, and Bruce Western. 2000. “Round Up the Usual Suspects: The Decline of

Unions in the Private Sector, 1973-1998.” Princeton University Working Paper 437.

Field, Alexander. 2003. “The Most Technologically Progressive Decade of the Century.” The

American Economic Review, 93(4): 1399–1413.

Field, Alexander. 2011. A Great Leap Forward: 1930s Depression and U.S. Economic Growth.

Yale University Press.

29

Goldin, Claudia, and Lawrence Katz. 2010. The Race between Education and Technology.

Harvard University Press.

Gordon, Robert. 1992. “Forward into the past: productivity retrogression in the electric gen-

erating industry.” NBER Working Paper 3988.

Gordon, Robert. 1999. “US economic growth since 1870: one big wave?” The American

Economic Review, 89(2): 123–128.

Gray, Rowena. 2013. “Taking technology to task: The skill content of technological change in

early twentieth century United States.” Explorations in Economic History, 50(3): 351–267.

Holmes, Thomas, and John Stevens. 2004. “Spatial distribution of economic activities in

North America.” Handbook of regional and urban economics, 4: 2797–2843.

Hughes, Tomas. 1993. Networks of Power: Electrification in Western Society, 1880-1930. John

Hopkins University Press.

Jerome, Harry. 1934. Mechanization in industry. National Bureau of Economic Research, New

York.

Jovanovic, Boyan, and Peter L Rousseau. 2005. “General Purpose Technologies.” In Hand-

book of Economic Growth. , ed. Philippe Aghion and Steven N Durlauf, 1181–1224. Elsevier.

Kendrick, John. 1961. Productivity trends in the United States. National Bureau of Economic

Research, New York.

Keynes, John Maynard. 1963. “Economic Possibilities for our Grandchildren.” 358–373. W.

W. Norton and Company.

Kimbrough, Gray, and Kenneth Snowden. 2007. “The Spatial Character of Housing De-

pression in the 1930s.” 2007 Economic History Association Meetings Papers.

Morin, Miguel Barroso. 2014. “General Purpose Technologies: engines of change?” PhD diss.,

Columbia University.

Orchard, Dennis. 1962. Concrete Technology. Vol. 2, John Wiley and Sons.

30

Sleight, Reuben. 1930. The 1930 award of the American Superpower Corporation in connection

with the Bonbright prize essay contest of 1925. American Superpower Corporation, Dover,

Delaware.

Stigler, George, and Claire Friedland. 1962. “What Can Regulators Regulate? The Case

of Electricity.” Journal of Law and Economics, 5: 1–16.

Syverson, Chad. 2004. “Market Structure and Productivity: A Concrete Example.” Journal

of Political Economy, 112(6): 1181–1222.

Syverson, Chad. 2013. “Will history repeat itself? Comments.” International Productivity

Monitor, 25: 37–40.

Woirol, Gregory. 1996. The Technological Unemployment and Structural Unemployment De-

bates. Greenwood Press, Westport, Connecticut.

Woolf, Arthur. 1984. “Electricity, productivity, and labor saving: American manufacturing,

1900-1929.” Explorations in Economic History, 21(2): 176–191.

Ziebarth, Nicolas. 2011. “Misallocation and Productivity during the Great Depression.” North-

western University manuscript (accessed 14 February 2012).

Data sources

Bureau of the Census. 1974. “Historical statistics of the United States: colonial times

to 1970.” Government Printing Office, Washington DC.

Bureau of the Census. Various issues. “Census of Electric Light and Power Stations.”

Government Printing Office, Washington DC.

Federal Power Commission. 1937. “Domestic and Residential Rates in Effect January

1, 1936 with Trends in Residential Rates from 1924 to 1936, Cities of 50,000 Population

and Over,” Government Printing Office, Washington DC.

31

Hay, Duncan. 1991. “Hydroelectric development in the United States, 1880-1940.”

Edison Electrical Institute, Washington DC.

McGraw-Hill Catalog and Directory Company. 1928. “McGraw Central Station

Directory,” New York.

National Electric Light Association. 1929. “Electrical research statistics,” New York.

National Electric Light Association. 1931. “The electric light and power industry in

the United States,” New York.

Tennessee Valley Authority. 1947. “Concrete production and control,” United States

Government Printing Office.

U.S. Bureau of Reclamation. 2011. “Hydropower Resource Assessment at Existing

Reclamation Facilities”

32

A Census of Manufactures for the concrete industry

A.1 Matching across years

I matched plants between years 1929 and 1935 according to a similar procedure as Bres-

nahan and Raff (1991). Some plants sent two schedules to the Census Bureau, such as one

by the plant and another by the general office; on two occasions, I aggregated them into a

new plant by either averaging their responses if the two schedules covered the same period

of operation, or by summing their results if they covered different periods. I considered

that two plants were a match if:

1. one plant is from 1929 and the other from 1935,

2. the two plants are located in the same state, county, and city,

3. one of the following conditions hold:

(a) the name fields coincide (name of plant, name of owner, or their change) and

the location fields coincide (same street location in both years, or the street

location in a year coincides with the general office location in another year),

(b) the name fields coincide, one of the plants did not report a street location, and

they are the only plants in that state, county, and city,

4. no other plants match criteria (1-3).

33

As an example of condition 2, I considered small cities included in larger cities to be the

same, such as Flushing and New York. I also considered nearby cities to be the same,

such as Edina and Minneapolis, since concrete plants sometimes reported the location of

the plant and sometimes the post office address of the general office. As an example of

condition 3 (a), it is verified between a plant in 1935 with name “Gehirs” and address “23

Conklin St,” and a plant in 1929 with owner “Gehirs” and address “Conklin street and

Liberty Avenue.” As an example of condition 4, if two plants in Rockford, Illinois, share

the name “Rockford plant” in 1929, then none is matched to the “Rockford plant” in 1935.

This procedure produces 629 plants merged between 1929 and 1935. Out of the 2,435

concrete plants operating in 1929, more than two thirds exited the market; out of the

1,108 concrete plants operating in 1935, a third entered the market.

The schedules changed slightly across plants. Some concrete plants in 1929 filled a sched-

ule for the Census of Mines and Quarries, which omitted questions about electricity con-

sumption and the quantity of output. Some plants filled other schedules and reported

their output in different units, e.g. the number of laundry trays instead of their weight.

34

A.2 Data for the Census of Manufactures in other years

The schedules before 1929 and after 1935 were lost. The Census Bureau used them to

compile information for the Statistical Abstracts and publications of the manufacturing

industry. After such compilation, an Act of Congress gave the right to destroy the sched-

ules. A 1971 letter by Dennis Rousey, Acting Chief of the Industrial and Social Branch,

mentioned that “Since 1900, the schedules of agriculture censuses have been disposed of

under Congressional authorization,” with the manufacturing schedules possibly having a

similar fate. An archivist told me that he was surprised that the schedules for 1929 to

1935 even survived, which he attributed to the relevance of the economic downturn. I

searched for earlier or later schedules extensively and found only one surviving schedule

from 1925, for the Crow Indian Mill in Colorado and kept at the National Archives in

Denver, and one surviving schedule from 1939, for a German-owned company and the

German American Bund that was seized during World War II. The schedules for the 1947

Census of Manufactures were transferred to non-safety microfilm, are disintegrating, and

are “unavailable to researchers [because of] preservation issues and concerns.”15

A.3 Categories of employment

The Census asked about two categories of employment, wage-earners and salaried workers,

described in detail below. Wage-earners are present in all years and represent around

90% of employment. Officers of the corporation were sometimes reported on a special

administrative schedule that is absent from the Census of Manufactures. In 1929, the

Census seems to have included engineers and other technical employees as wage-earners.

In 1935, technical employees had a separate category. This chapter considers all categories

of employment, excluding proprietors, who had no salary, and salaried officers of the

corporation, who were sometimes reported on a different form. The details of employment

categories suggest that the two types of employment are different from skilled/unskilled

and from routine/nonroutine occupations.15Electronic correspondence with the National Archives at College Park, Maryland.

35

• Categories of employment in 1929:

– Proprietor or firm members

– Principal officers of corporations

– “Managers, superintendents, and other responsible administrative employees;

foremen and overseers who devote all or the greater part of their time to super-

visory duties; clerks, stenographers, bookkeepers, and other clerical employees

on salary.”

– Wage-earners: “Skilled and unskilled workers of all classes, including engineers,

firemen, watchmen, packers; also foremen and overseers in minor positions who

perform work similar to that done by the employees under their supervision.”

• Categories of employment in 1935:

– Proprietor or firm members

– Salaried officers of the corporation

– Supervisory employees: “managers, superintendents, and other responsible ad-

ministrative employees (including plant foremen whose duties are primarily

supervisory but not including foremen and overseers in minor positions who

perform work similar to that of the employees under their supervision”

– Technical employees: “trained technicians, such as chemists, electrical and me-

chanical engineers, designers, who hold responsible positions requiring technical

training and whose supervisory duties, if any, are only incidental”

– Clerical employees: “clerks, stenographers, bookkeepers, timekeepers, and other

clerical employees (including laboratory assistants, draftsmen), whether in the

office or in the factory”

36

– Wage-earners: “all time and piece workers employed in the plant (including the

power plant and the maintenance, shipping, warehousing, and other depart-

ments) covered by this report, not including employees reported above. Include

here working foremen and gang and straw bosses, but nor foremen whose duties

are primarily supervisory.”

A.4 Measurement of plant-level variables and industry background

The histograms in Figure 8 suggest that the labor share of income have bell-shaped

frequency curves with accurate measurement. The Census Bureau checked thoroughly

these variables and mailed the plant for more information when it found outliers. In

contrast, Figure 9 suggests that the average price of electricity has considerable variation,

up to 1 dollar per kilowatt-hour, at a time when the average price for the United States

was 2.6 cents per kilowatt-hour.

05

1015

Perc

ent

0 .2 .4 .6 .8Labor share of income in 1929

05

1015

Perc

ent

0 .2 .4 .6 .8 1Labor share of income in 1935

Figure 8: The labor share of revenue of concrete plants in 1929 and 1935 has a bell-shapeddistribution.

37

020

4060

80Pe

rcen

t

0 .5 1 1.5Plant-level price of electricity in 1929

020

4060

Perc

ent

0 .2 .4 .6 .8 1Plant-level price of electricity in 1935

Figure 9: The average price of electricity of concrete plants in 1929 and 1935 has a fat-taileddistribution.

This chapter considers the income pi,tYi,t to be revenue instead of value added. Revenue

is a more robust measure and contains fewer outliers: for example, some plants during the

Depression were operating at a loss and had negative value added (see Berman, Bound

and Griliches (1994, page 383) for a similar approach).

Tennessee Valley Authority (1947) details the production of concrete for the Tennessee

Valley Authority projects. It consists of mixing cement (often portland cement) with

water and an aggregate (crushed stone, sand, or gravel). Production of concrete starts

with collecting the aggregate, for example the sand of a river or the stone from a quarry. If

concrete plants own a quarry (in 1929, many plants were included in the Census of Mines

and Quarries rather than the Census of Manufactures), they may crush the stone to obtain

a finer aggregate. Otherwise, plants may buy the aggregate already crushed. Plants mix

the ingredients—cement, the aggregate, and water—to obtain a fluid substance that they

pour onto a mold. The substance hardens with time. Plants sometimes vibrate the mold

to achieve a more compact product. They cure the concrete product with water, as cement

requires a moist environment to harden further and increase strength. Plants may also

polish the concrete product with sandblasting—a jet of water mixed with sand under high

pressure to remove superficial irregularities. If plants convey the concrete product over a

long distance to the delivery location, the product bears the risk of un-mixing.

38

B Proofs

Proof of equations (2) and (3). This proof omits the plant index i. The firm maximizes

intertemporal profits

1X

t=0

D0,t

⇣Ai,tK

↵NE,i,tL

�NR,t

�K⇢

E,i,t + L⇢R,i,t

� �⇢ � wt (LNR,t + LR,t)� rNE,tKNE,t � rE,tKE,t

⌘,

where ⇢ = (� � 1) /�. The firm has no accumulation constraints on capital or labor and

the intertemporal maximization problem collapses to a sequence of static maximization

problems. The first-order conditions for profit-maximization, taking prices as given, are:

MPKNE,t =↵ Yt

KNE,t= rNE,t,

MPLNR,t =� Yt

LNR,t= wt,

MPKE,t = � Yt L⇢�1R,t

�K⇢

EI,t + L⇢R,t

��1= rE,t,

MPLR,t = � Yt L⇢�1R,t

�K⇢

E,t + L⇢R,t

��1= wt,

where MPF is the marginal product of factor F . The ratio of electric capital to employ-

ment in routine occupations is:

KE,t

LR,t=

✓rE,t

wt

◆��

.

39

The labor share of income is:

wtLt

Yt=

wtLNR,t

Yt+

wtLR,t

Yt,

= � + �L⇢R,t

�K⇢

E,t + L⇢R,t

��1,

= � + �

✓1 +

✓KE,t

LR,t

◆⇢◆�1

,

= � + �

1 +

✓rE,t

wt

◆�⇢�!�1

,

= � + �

1 +

✓rE,t

wt

◆1��!�1

.

The routine share of labor is:

LR,t

Lt=

LR,t

LNR,t + LR,t,

=

✓1 +

LNR,t

LR,t

◆�1

,

=

1 +

�

�

r1��E,t + w1��

t

w1��t

!�1

.

The electric capital-total labor ratio is :

KE,t

Lt=

KE,t

LR,t

LR,t

Lt,

=

✓rE,t

wt

◆�� 1 +

�

�

+

�

�

✓rE,t

wt

◆1��!�1

,

=

✓rE,t

wt

◆�1 �

�

+

✓1 +

�

�

◆✓rE,t

wt

◆��1!�1

.

In short, the labor share of income wtLt/ptYt is increasing in rE,t, electrical intensity

KE,t/Lt is decreasing in rE,t, and labor productivity Yt/Lt is decreasing in rE,t.

40

C Electricity data and background

C.1 Other measures of the price of electricity

Other measures of the price of electricity exist during this period but they are inferior

to the state-level price of electricity used in the baseline regressions. First, the price

of electricity paid by ice plants (Ziebarth, 2011) covers cities that coincide with only

200 concrete plants. Second, the city-level price of electricity for residential consumers

for a typical bill of 25, 100, or 250 kilowatt-hours (Federal Power Commission, 1937) is

a survey with measurement error due to retrospective questions asked in 1936, concerns

residential consumers instead of industrial consumers, and has a significantly lower amount

than the average demand by concrete plants in 1929 (1400 kilowatt-hours per month for

concrete plants versus 250 kilowatt-hours for residential consumers), and they are also on

different rate schedules (see the next section). Third, the price of electricity by municipal

utilities from the Census of Electric Light and Power Stations in 1927 and 1937 concerns

a small market (5% of total kilowatt-hours).16 Fourth, the Census of Electric Light and

Power Stations published the price of electricity from both public and private utilities to

industrial consumers, split by “small” (retail) and “large” (wholesale), but the “wholesale”

numbers exist only half of the states to prevent disclosure of establishment information.

To the best of my knowledge, there are no other measures for the price of electricity that

are disaggregated geographically over this period.

To show how the price of electricity at the plant-level is plagued with fixed costs, figure 10

shows a scatter plot of the change in the state-level price of electricity and a Paasche index

of the change in the price of electricity at the plant-level aggregated at the state-level:

the two measures should be positively related but are negatively related.

16Census of Electric Light and Power Stations, 1927, page 71.

41

ME

NH

MA

NY

NJ

PAOHIN ILMI

WIMN

IA

MO

ND

SD

NE

KSMD

DCVA

WVNC

FL

KY

TNAR

LA

TXCO

WA

OR

CA-3

-2-1

01

2Ch

ange

in P

aasc

he in

dex,

29-

35

-.6 -.4 -.2 0 .2Change in state-level price of electricity, 27-37

Slope: -2.38 t-statistic: -2.26R2: 0.14 Correlation: -38% Observations: 33

Figure 10: The change in the Paasche index of the price of electricity is negatively related to thechange in the state-level price of electricity.

C.2 Pricing of electricity

Electric utilities offered many rate schedules, detailed by the Federal Power Commission

in a published glossary in 1936. All rates have a component of capacity, in kilowatts or

horsepower, and of energy, in kilowatt-hours or Joules.

42

An electric bill consists of three types of charges: a customer charge, a demand charge, and

an energy charge. The Federal Power Commission defines “customer charge” or “service

charge” as “a component part of a rate schedule providing that a customer must pay

a certain definite sum in a specified period (usually 1 month) without regard to the

consumption of energy or the demand, for which he can use no energy or demand.”

It defines a “demand charge” as “a component part of a rate schedule which provides

for a charge based upon the customer’s demand or equivalent, without regard to the

consumption of energy.” It defines “energy charge” as “a component part of a rate schedule

that provides for a charge based upon the amount of energy consumed.” In short, the

customer pays a service charge for connecting to the grid, a demand charge for the right

to use a given capacity from the grid, and an energy charge for consumption of electricity.

Most rate schedules also define “maximum demand,” which is often the aggregate capacity

of electric appliances commonly used. For example, a plant may have a primary motor

and a stand-by motor, each with a capacity of 100 kW. The plant may normally use only

the primary motor and contracts for a maximum demand of 100 kW. If the plant happens

to use both motors at the same time, it will have to pay a higher price for using more

capacity than the maximum demand.

Electric utilities offered up to eight different schedules. The flat rate schedule “provides

for a specified charge per unit of time, irrespective of the amount of electric energy taken.

For example: $2 per month per customer up to and including 6-50 watt lamps.”

The straight line meter rate schedule “provides for a constant charge per unit of energy

regardless of the amount consumed. For example: 5 cents per kilowatt-hour.”

The flat demand rate schedule “bases the billing either on the demand or on some fixed

characteristic indicative of demand but provides no charge for energy. For example: $50.00

per year per horsepower of demand.”

43

The flat and meter rate schedule is a two-part tariff with “two components, the first of

which is a customer (or service) charge and the second of which is a price for the energy

consumed.”

The block meter rate schedule “divides the total amount of energy to be consumed during

a definite period into prescribed blocks and provides a different rate for each.”

The Hopkinson demand rate schedule has “two components, the first of which is a charge

for demand, and the second a charge for the energy consumed.”

The block Hopkinson demand rate schedule has “either the demand charge or the energy

charge or both are arranged in blocks. For example, a demand charge of $1.25 for the

first 50 kilowatts of maximum demand per month, and $1.00 per kilowatt for all above 50

kilowatts of maximum demand per month. Plus: an energy charge of 3 cents per kilowatt-

hour for the first 1,000 kilowatt-hours used per month, and 1 cent per kilowatt-hour for

all energy used in excess of 1,000 kilowatt-hours per month.”

The step meter rate schedule has “a charge per unit of energy [that] is constant for all

kilowatt-hours consumed during the billing period, the charge per unit depending upon

the total consumption. For example: if 1 to 25 kilowatt-hours are used in a month, 5 cents

per kilowatt-hour; if 26 to 50 kilowatt-hours are used in a month, 3 cents per kilowatt-hour

(for all the energy including the first 25 kilowatt-hours).”

The three-part rate schedule “provides three components for determining the total bill:

customer charge, demand charge, and energy charge. For example: 50 cents per month

per meter. Plus: a demand charge of $1.25 per month per kilowatt for the first 25

kilowatts of maximum demand in the month; 90 cents per month per kilowatt for the

excess of the maximum demand over 25 kilowatts. Plus: an energy charge of 1.5 cents

per kilowatt-hour.”

Furthermore, rate schedules may have clauses providing for additional charges in the event

of large increases in the price of coal, the price of commodities, or wages.

44

C.3 Technical progress in the generation of electricity

Figure 11 illustrates the exponential decrease in the price of electricity over the first half

of the 20th century. Gordon (1992, Table 1) estimates the rate of decrease in the price of

electricity at 7% per year between 1899 and 1948. The real price of electricity increased

slightly during the Great Depression because of deflation in the consumer price index. In

a more general model with irreversible investment, firms would have difficulty adjusting

their capital stock to cyclical changes in the price of electricity and would react to the

trend in the price of electricity rather than to the fluctuations. Furthermore, the nominal

price of electricity decreased by 0.02 log-points in the sample of concrete plants (see Table

1).

15

1015

Pric

e of

ele

ctric

ity (1

950

= 1)

1900 1910 1920 1930 1940 1950Year

Residential consumers Industrial consumersAll consumers

Figure 11: The real price of electricity decreased exponentially in the first half of the 20th century.The price of electricity is in cents per kilowatt-hour from the Historical Statistics of the United States, series Db234,Db235, and Db237. The price deflator is the consumer price index from the BLS, series Cc1. The rate of decreaseof the price of electricity for residential consumers is 5.8%.

The technology to produce electricity from coal improved over the first half of the 20th

century, but hydroelectric technology did not:

45

In generating electricity from coal even the largest and most modern electric

power stations are able to utilize only about 25 per cent of the heat units

available in the coal. ... On the other hand, modern hydro-electric machin-

ery now transforms into electricity more than 90 per cent of the energy in

falling waters, leaving little opportunity for radical improvements in present-

day hydro-electric practice. (The electric light and power industry, 1931, page

43)

Hughes (1993) describes another source of technical progress with economies of scale. The

interested reader is referred to his account of electrification in Western Society over the

period 1880-1930.

46

the labor market consequences of electricity adoption in...

Documents